src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEw_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_single.h"
  50 #include "kernelutil_x86_sse2_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_sse2_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_sse2_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  96     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  97     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     __m128i          ewitab;
 109     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     real             *ewtab;
 111     __m128           dummy_mask,cutoff_mask;
 112     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 113     __m128           one     = _mm_set1_ps(1.0);
 114     __m128           two     = _mm_set1_ps(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_ps(fr->epsfac);
 127     charge           = mdatoms->chargeA;
 128
 129     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 130     ewtab            = fr->ic->tabq_coul_FDV0;
 131     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 132     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 133
 134     /* Setup water-specific parameters */
 135     inr              = nlist->iinr[0];
 136     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 137     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 138     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 139
 140     jq1              = _mm_set1_ps(charge[inr+1]);
 141     jq2              = _mm_set1_ps(charge[inr+2]);
 142     jq3              = _mm_set1_ps(charge[inr+3]);
 143     qq11             = _mm_mul_ps(iq1,jq1);
 144     qq12             = _mm_mul_ps(iq1,jq2);
 145     qq13             = _mm_mul_ps(iq1,jq3);
 146     qq21             = _mm_mul_ps(iq2,jq1);
 147     qq22             = _mm_mul_ps(iq2,jq2);
 148     qq23             = _mm_mul_ps(iq2,jq3);
 149     qq31             = _mm_mul_ps(iq3,jq1);
 150     qq32             = _mm_mul_ps(iq3,jq2);
 151     qq33             = _mm_mul_ps(iq3,jq3);
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = jnrC = jnrD = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157     j_coord_offsetC = 0;
 158     j_coord_offsetD = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     for(iidx=0;iidx<4*DIM;iidx++)
 164     {
 165         scratch[iidx] = 0.0;
 166     }
 167
 168     /* Start outer loop over neighborlists */
 169     for(iidx=0; iidx<nri; iidx++)
 170     {
 171         /* Load shift vector for this list */
 172         i_shift_offset   = DIM*shiftidx[iidx];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 184                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 185
 186         fix1             = _mm_setzero_ps();
 187         fiy1             = _mm_setzero_ps();
 188         fiz1             = _mm_setzero_ps();
 189         fix2             = _mm_setzero_ps();
 190         fiy2             = _mm_setzero_ps();
 191         fiz2             = _mm_setzero_ps();
 192         fix3             = _mm_setzero_ps();
 193         fiy3             = _mm_setzero_ps();
 194         fiz3             = _mm_setzero_ps();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm_setzero_ps();
 198
 199         /* Start inner kernel loop */
 200         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 201         {
 202
 203             /* Get j neighbor index, and coordinate index */
 204             jnrA             = jjnr[jidx];
 205             jnrB             = jjnr[jidx+1];
 206             jnrC             = jjnr[jidx+2];
 207             jnrD             = jjnr[jidx+3];
 208             j_coord_offsetA  = DIM*jnrA;
 209             j_coord_offsetB  = DIM*jnrB;
 210             j_coord_offsetC  = DIM*jnrC;
 211             j_coord_offsetD  = DIM*jnrD;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 215                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 216                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 217
 218             /* Calculate displacement vector */
 219             dx11             = _mm_sub_ps(ix1,jx1);
 220             dy11             = _mm_sub_ps(iy1,jy1);
 221             dz11             = _mm_sub_ps(iz1,jz1);
 222             dx12             = _mm_sub_ps(ix1,jx2);
 223             dy12             = _mm_sub_ps(iy1,jy2);
 224             dz12             = _mm_sub_ps(iz1,jz2);
 225             dx13             = _mm_sub_ps(ix1,jx3);
 226             dy13             = _mm_sub_ps(iy1,jy3);
 227             dz13             = _mm_sub_ps(iz1,jz3);
 228             dx21             = _mm_sub_ps(ix2,jx1);
 229             dy21             = _mm_sub_ps(iy2,jy1);
 230             dz21             = _mm_sub_ps(iz2,jz1);
 231             dx22             = _mm_sub_ps(ix2,jx2);
 232             dy22             = _mm_sub_ps(iy2,jy2);
 233             dz22             = _mm_sub_ps(iz2,jz2);
 234             dx23             = _mm_sub_ps(ix2,jx3);
 235             dy23             = _mm_sub_ps(iy2,jy3);
 236             dz23             = _mm_sub_ps(iz2,jz3);
 237             dx31             = _mm_sub_ps(ix3,jx1);
 238             dy31             = _mm_sub_ps(iy3,jy1);
 239             dz31             = _mm_sub_ps(iz3,jz1);
 240             dx32             = _mm_sub_ps(ix3,jx2);
 241             dy32             = _mm_sub_ps(iy3,jy2);
 242             dz32             = _mm_sub_ps(iz3,jz2);
 243             dx33             = _mm_sub_ps(ix3,jx3);
 244             dy33             = _mm_sub_ps(iy3,jy3);
 245             dz33             = _mm_sub_ps(iz3,jz3);
 246
 247             /* Calculate squared distance and things based on it */
 248             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 249             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 250             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 251             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 252             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 253             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 254             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 255             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 256             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 257
 258             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 259             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 260             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 261             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 262             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 263             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 264             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 265             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 266             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 267
 268             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 269             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 270             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 271             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 272             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 273             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 274             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 275             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 276             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 277
 278             fjx1             = _mm_setzero_ps();
 279             fjy1             = _mm_setzero_ps();
 280             fjz1             = _mm_setzero_ps();
 281             fjx2             = _mm_setzero_ps();
 282             fjy2             = _mm_setzero_ps();
 283             fjz2             = _mm_setzero_ps();
 284             fjx3             = _mm_setzero_ps();
 285             fjy3             = _mm_setzero_ps();
 286             fjz3             = _mm_setzero_ps();
 287
 288             /**************************
 289              * CALCULATE INTERACTIONS *
 290              **************************/
 291
 292             r11              = _mm_mul_ps(rsq11,rinv11);
 293
 294             /* EWALD ELECTROSTATICS */
 295
 296             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 297             ewrt             = _mm_mul_ps(r11,ewtabscale);
 298             ewitab           = _mm_cvttps_epi32(ewrt);
 299             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 300             ewitab           = _mm_slli_epi32(ewitab,2);
 301             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 302             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 303             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 304             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 305             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 306             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 307             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 308             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 309             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 310
 311             /* Update potential sum for this i atom from the interaction with this j atom. */
 312             velecsum         = _mm_add_ps(velecsum,velec);
 313
 314             fscal            = felec;
 315
 316             /* Calculate temporary vectorial force */
 317             tx               = _mm_mul_ps(fscal,dx11);
 318             ty               = _mm_mul_ps(fscal,dy11);
 319             tz               = _mm_mul_ps(fscal,dz11);
 320
 321             /* Update vectorial force */
 322             fix1             = _mm_add_ps(fix1,tx);
 323             fiy1             = _mm_add_ps(fiy1,ty);
 324             fiz1             = _mm_add_ps(fiz1,tz);
 325
 326             fjx1             = _mm_add_ps(fjx1,tx);
 327             fjy1             = _mm_add_ps(fjy1,ty);
 328             fjz1             = _mm_add_ps(fjz1,tz);
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             r12              = _mm_mul_ps(rsq12,rinv12);
 335
 336             /* EWALD ELECTROSTATICS */
 337
 338             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 339             ewrt             = _mm_mul_ps(r12,ewtabscale);
 340             ewitab           = _mm_cvttps_epi32(ewrt);
 341             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 342             ewitab           = _mm_slli_epi32(ewitab,2);
 343             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 344             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 345             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 346             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 347             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 348             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 349             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 350             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 351             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velecsum         = _mm_add_ps(velecsum,velec);
 355
 356             fscal            = felec;
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = _mm_mul_ps(fscal,dx12);
 360             ty               = _mm_mul_ps(fscal,dy12);
 361             tz               = _mm_mul_ps(fscal,dz12);
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_add_ps(fix1,tx);
 365             fiy1             = _mm_add_ps(fiy1,ty);
 366             fiz1             = _mm_add_ps(fiz1,tz);
 367
 368             fjx2             = _mm_add_ps(fjx2,tx);
 369             fjy2             = _mm_add_ps(fjy2,ty);
 370             fjz2             = _mm_add_ps(fjz2,tz);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             r13              = _mm_mul_ps(rsq13,rinv13);
 377
 378             /* EWALD ELECTROSTATICS */
 379
 380             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 381             ewrt             = _mm_mul_ps(r13,ewtabscale);
 382             ewitab           = _mm_cvttps_epi32(ewrt);
 383             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 384             ewitab           = _mm_slli_epi32(ewitab,2);
 385             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 386             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 387             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 388             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 389             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 390             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 391             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 392             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 393             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velecsum         = _mm_add_ps(velecsum,velec);
 397
 398             fscal            = felec;
 399
 400             /* Calculate temporary vectorial force */
 401             tx               = _mm_mul_ps(fscal,dx13);
 402             ty               = _mm_mul_ps(fscal,dy13);
 403             tz               = _mm_mul_ps(fscal,dz13);
 404
 405             /* Update vectorial force */
 406             fix1             = _mm_add_ps(fix1,tx);
 407             fiy1             = _mm_add_ps(fiy1,ty);
 408             fiz1             = _mm_add_ps(fiz1,tz);
 409
 410             fjx3             = _mm_add_ps(fjx3,tx);
 411             fjy3             = _mm_add_ps(fjy3,ty);
 412             fjz3             = _mm_add_ps(fjz3,tz);
 413
 414             /**************************
 415              * CALCULATE INTERACTIONS *
 416              **************************/
 417
 418             r21              = _mm_mul_ps(rsq21,rinv21);
 419
 420             /* EWALD ELECTROSTATICS */
 421
 422             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 423             ewrt             = _mm_mul_ps(r21,ewtabscale);
 424             ewitab           = _mm_cvttps_epi32(ewrt);
 425             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 426             ewitab           = _mm_slli_epi32(ewitab,2);
 427             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 428             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 429             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 430             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 431             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 432             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 433             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 434             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
 435             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velecsum         = _mm_add_ps(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             /* Calculate temporary vectorial force */
 443             tx               = _mm_mul_ps(fscal,dx21);
 444             ty               = _mm_mul_ps(fscal,dy21);
 445             tz               = _mm_mul_ps(fscal,dz21);
 446
 447             /* Update vectorial force */
 448             fix2             = _mm_add_ps(fix2,tx);
 449             fiy2             = _mm_add_ps(fiy2,ty);
 450             fiz2             = _mm_add_ps(fiz2,tz);
 451
 452             fjx1             = _mm_add_ps(fjx1,tx);
 453             fjy1             = _mm_add_ps(fjy1,ty);
 454             fjz1             = _mm_add_ps(fjz1,tz);
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             r22              = _mm_mul_ps(rsq22,rinv22);
 461
 462             /* EWALD ELECTROSTATICS */
 463
 464             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 465             ewrt             = _mm_mul_ps(r22,ewtabscale);
 466             ewitab           = _mm_cvttps_epi32(ewrt);
 467             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 468             ewitab           = _mm_slli_epi32(ewitab,2);
 469             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 470             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 471             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 472             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 473             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 474             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 475             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 476             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
 477             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 478
 479             /* Update potential sum for this i atom from the interaction with this j atom. */
 480             velecsum         = _mm_add_ps(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             /* Calculate temporary vectorial force */
 485             tx               = _mm_mul_ps(fscal,dx22);
 486             ty               = _mm_mul_ps(fscal,dy22);
 487             tz               = _mm_mul_ps(fscal,dz22);
 488
 489             /* Update vectorial force */
 490             fix2             = _mm_add_ps(fix2,tx);
 491             fiy2             = _mm_add_ps(fiy2,ty);
 492             fiz2             = _mm_add_ps(fiz2,tz);
 493
 494             fjx2             = _mm_add_ps(fjx2,tx);
 495             fjy2             = _mm_add_ps(fjy2,ty);
 496             fjz2             = _mm_add_ps(fjz2,tz);
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             r23              = _mm_mul_ps(rsq23,rinv23);
 503
 504             /* EWALD ELECTROSTATICS */
 505
 506             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 507             ewrt             = _mm_mul_ps(r23,ewtabscale);
 508             ewitab           = _mm_cvttps_epi32(ewrt);
 509             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 510             ewitab           = _mm_slli_epi32(ewitab,2);
 511             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 512             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 513             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 514             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 515             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 516             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 517             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 518             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
 519             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
 520
 521             /* Update potential sum for this i atom from the interaction with this j atom. */
 522             velecsum         = _mm_add_ps(velecsum,velec);
 523
 524             fscal            = felec;
 525
 526             /* Calculate temporary vectorial force */
 527             tx               = _mm_mul_ps(fscal,dx23);
 528             ty               = _mm_mul_ps(fscal,dy23);
 529             tz               = _mm_mul_ps(fscal,dz23);
 530
 531             /* Update vectorial force */
 532             fix2             = _mm_add_ps(fix2,tx);
 533             fiy2             = _mm_add_ps(fiy2,ty);
 534             fiz2             = _mm_add_ps(fiz2,tz);
 535
 536             fjx3             = _mm_add_ps(fjx3,tx);
 537             fjy3             = _mm_add_ps(fjy3,ty);
 538             fjz3             = _mm_add_ps(fjz3,tz);
 539
 540             /**************************
 541              * CALCULATE INTERACTIONS *
 542              **************************/
 543
 544             r31              = _mm_mul_ps(rsq31,rinv31);
 545
 546             /* EWALD ELECTROSTATICS */
 547
 548             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 549             ewrt             = _mm_mul_ps(r31,ewtabscale);
 550             ewitab           = _mm_cvttps_epi32(ewrt);
 551             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 552             ewitab           = _mm_slli_epi32(ewitab,2);
 553             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 554             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 555             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 556             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 557             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 558             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 559             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 560             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
 561             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
 562
 563             /* Update potential sum for this i atom from the interaction with this j atom. */
 564             velecsum         = _mm_add_ps(velecsum,velec);
 565
 566             fscal            = felec;
 567
 568             /* Calculate temporary vectorial force */
 569             tx               = _mm_mul_ps(fscal,dx31);
 570             ty               = _mm_mul_ps(fscal,dy31);
 571             tz               = _mm_mul_ps(fscal,dz31);
 572
 573             /* Update vectorial force */
 574             fix3             = _mm_add_ps(fix3,tx);
 575             fiy3             = _mm_add_ps(fiy3,ty);
 576             fiz3             = _mm_add_ps(fiz3,tz);
 577
 578             fjx1             = _mm_add_ps(fjx1,tx);
 579             fjy1             = _mm_add_ps(fjy1,ty);
 580             fjz1             = _mm_add_ps(fjz1,tz);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             r32              = _mm_mul_ps(rsq32,rinv32);
 587
 588             /* EWALD ELECTROSTATICS */
 589
 590             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 591             ewrt             = _mm_mul_ps(r32,ewtabscale);
 592             ewitab           = _mm_cvttps_epi32(ewrt);
 593             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 594             ewitab           = _mm_slli_epi32(ewitab,2);
 595             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 596             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 597             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 598             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 599             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 600             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 601             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 602             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
 603             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
 604
 605             /* Update potential sum for this i atom from the interaction with this j atom. */
 606             velecsum         = _mm_add_ps(velecsum,velec);
 607
 608             fscal            = felec;
 609
 610             /* Calculate temporary vectorial force */
 611             tx               = _mm_mul_ps(fscal,dx32);
 612             ty               = _mm_mul_ps(fscal,dy32);
 613             tz               = _mm_mul_ps(fscal,dz32);
 614
 615             /* Update vectorial force */
 616             fix3             = _mm_add_ps(fix3,tx);
 617             fiy3             = _mm_add_ps(fiy3,ty);
 618             fiz3             = _mm_add_ps(fiz3,tz);
 619
 620             fjx2             = _mm_add_ps(fjx2,tx);
 621             fjy2             = _mm_add_ps(fjy2,ty);
 622             fjz2             = _mm_add_ps(fjz2,tz);
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             r33              = _mm_mul_ps(rsq33,rinv33);
 629
 630             /* EWALD ELECTROSTATICS */
 631
 632             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 633             ewrt             = _mm_mul_ps(r33,ewtabscale);
 634             ewitab           = _mm_cvttps_epi32(ewrt);
 635             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 636             ewitab           = _mm_slli_epi32(ewitab,2);
 637             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 638             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 639             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 640             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 641             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 642             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 643             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 644             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
 645             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
 646
 647             /* Update potential sum for this i atom from the interaction with this j atom. */
 648             velecsum         = _mm_add_ps(velecsum,velec);
 649
 650             fscal            = felec;
 651
 652             /* Calculate temporary vectorial force */
 653             tx               = _mm_mul_ps(fscal,dx33);
 654             ty               = _mm_mul_ps(fscal,dy33);
 655             tz               = _mm_mul_ps(fscal,dz33);
 656
 657             /* Update vectorial force */
 658             fix3             = _mm_add_ps(fix3,tx);
 659             fiy3             = _mm_add_ps(fiy3,ty);
 660             fiz3             = _mm_add_ps(fiz3,tz);
 661
 662             fjx3             = _mm_add_ps(fjx3,tx);
 663             fjy3             = _mm_add_ps(fjy3,ty);
 664             fjz3             = _mm_add_ps(fjz3,tz);
 665
 666             fjptrA             = f+j_coord_offsetA;
 667             fjptrB             = f+j_coord_offsetB;
 668             fjptrC             = f+j_coord_offsetC;
 669             fjptrD             = f+j_coord_offsetD;
 670
 671             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 672                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 673
 674             /* Inner loop uses 369 flops */
 675         }
 676
 677         if(jidx<j_index_end)
 678         {
 679
 680             /* Get j neighbor index, and coordinate index */
 681             jnrlistA         = jjnr[jidx];
 682             jnrlistB         = jjnr[jidx+1];
 683             jnrlistC         = jjnr[jidx+2];
 684             jnrlistD         = jjnr[jidx+3];
 685             /* Sign of each element will be negative for non-real atoms.
 686              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 687              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 688              */
 689             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 690             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 691             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 692             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 693             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 694             j_coord_offsetA  = DIM*jnrA;
 695             j_coord_offsetB  = DIM*jnrB;
 696             j_coord_offsetC  = DIM*jnrC;
 697             j_coord_offsetD  = DIM*jnrD;
 698
 699             /* load j atom coordinates */
 700             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 701                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 702                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 703
 704             /* Calculate displacement vector */
 705             dx11             = _mm_sub_ps(ix1,jx1);
 706             dy11             = _mm_sub_ps(iy1,jy1);
 707             dz11             = _mm_sub_ps(iz1,jz1);
 708             dx12             = _mm_sub_ps(ix1,jx2);
 709             dy12             = _mm_sub_ps(iy1,jy2);
 710             dz12             = _mm_sub_ps(iz1,jz2);
 711             dx13             = _mm_sub_ps(ix1,jx3);
 712             dy13             = _mm_sub_ps(iy1,jy3);
 713             dz13             = _mm_sub_ps(iz1,jz3);
 714             dx21             = _mm_sub_ps(ix2,jx1);
 715             dy21             = _mm_sub_ps(iy2,jy1);
 716             dz21             = _mm_sub_ps(iz2,jz1);
 717             dx22             = _mm_sub_ps(ix2,jx2);
 718             dy22             = _mm_sub_ps(iy2,jy2);
 719             dz22             = _mm_sub_ps(iz2,jz2);
 720             dx23             = _mm_sub_ps(ix2,jx3);
 721             dy23             = _mm_sub_ps(iy2,jy3);
 722             dz23             = _mm_sub_ps(iz2,jz3);
 723             dx31             = _mm_sub_ps(ix3,jx1);
 724             dy31             = _mm_sub_ps(iy3,jy1);
 725             dz31             = _mm_sub_ps(iz3,jz1);
 726             dx32             = _mm_sub_ps(ix3,jx2);
 727             dy32             = _mm_sub_ps(iy3,jy2);
 728             dz32             = _mm_sub_ps(iz3,jz2);
 729             dx33             = _mm_sub_ps(ix3,jx3);
 730             dy33             = _mm_sub_ps(iy3,jy3);
 731             dz33             = _mm_sub_ps(iz3,jz3);
 732
 733             /* Calculate squared distance and things based on it */
 734             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 735             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 736             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 737             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 738             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 739             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 740             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 741             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 742             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 743
 744             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 745             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 746             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 747             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 748             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 749             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 750             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 751             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 752             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 753
 754             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 755             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 756             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 757             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 758             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 759             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 760             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 761             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 762             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 763
 764             fjx1             = _mm_setzero_ps();
 765             fjy1             = _mm_setzero_ps();
 766             fjz1             = _mm_setzero_ps();
 767             fjx2             = _mm_setzero_ps();
 768             fjy2             = _mm_setzero_ps();
 769             fjz2             = _mm_setzero_ps();
 770             fjx3             = _mm_setzero_ps();
 771             fjy3             = _mm_setzero_ps();
 772             fjz3             = _mm_setzero_ps();
 773
 774             /**************************
 775              * CALCULATE INTERACTIONS *
 776              **************************/
 777
 778             r11              = _mm_mul_ps(rsq11,rinv11);
 779             r11              = _mm_andnot_ps(dummy_mask,r11);
 780
 781             /* EWALD ELECTROSTATICS */
 782
 783             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 784             ewrt             = _mm_mul_ps(r11,ewtabscale);
 785             ewitab           = _mm_cvttps_epi32(ewrt);
 786             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 787             ewitab           = _mm_slli_epi32(ewitab,2);
 788             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 789             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 790             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 791             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 792             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 793             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 794             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 795             velec            = _mm_mul_ps(qq11,_mm_sub_ps(rinv11,velec));
 796             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
 797
 798             /* Update potential sum for this i atom from the interaction with this j atom. */
 799             velec            = _mm_andnot_ps(dummy_mask,velec);
 800             velecsum         = _mm_add_ps(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 805
 806             /* Calculate temporary vectorial force */
 807             tx               = _mm_mul_ps(fscal,dx11);
 808             ty               = _mm_mul_ps(fscal,dy11);
 809             tz               = _mm_mul_ps(fscal,dz11);
 810
 811             /* Update vectorial force */
 812             fix1             = _mm_add_ps(fix1,tx);
 813             fiy1             = _mm_add_ps(fiy1,ty);
 814             fiz1             = _mm_add_ps(fiz1,tz);
 815
 816             fjx1             = _mm_add_ps(fjx1,tx);
 817             fjy1             = _mm_add_ps(fjy1,ty);
 818             fjz1             = _mm_add_ps(fjz1,tz);
 819
 820             /**************************
 821              * CALCULATE INTERACTIONS *
 822              **************************/
 823
 824             r12              = _mm_mul_ps(rsq12,rinv12);
 825             r12              = _mm_andnot_ps(dummy_mask,r12);
 826
 827             /* EWALD ELECTROSTATICS */
 828
 829             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 830             ewrt             = _mm_mul_ps(r12,ewtabscale);
 831             ewitab           = _mm_cvttps_epi32(ewrt);
 832             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 833             ewitab           = _mm_slli_epi32(ewitab,2);
 834             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 835             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 836             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 837             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 838             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 839             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 840             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 841             velec            = _mm_mul_ps(qq12,_mm_sub_ps(rinv12,velec));
 842             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
 843
 844             /* Update potential sum for this i atom from the interaction with this j atom. */
 845             velec            = _mm_andnot_ps(dummy_mask,velec);
 846             velecsum         = _mm_add_ps(velecsum,velec);
 847
 848             fscal            = felec;
 849
 850             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 851
 852             /* Calculate temporary vectorial force */
 853             tx               = _mm_mul_ps(fscal,dx12);
 854             ty               = _mm_mul_ps(fscal,dy12);
 855             tz               = _mm_mul_ps(fscal,dz12);
 856
 857             /* Update vectorial force */
 858             fix1             = _mm_add_ps(fix1,tx);
 859             fiy1             = _mm_add_ps(fiy1,ty);
 860             fiz1             = _mm_add_ps(fiz1,tz);
 861
 862             fjx2             = _mm_add_ps(fjx2,tx);
 863             fjy2             = _mm_add_ps(fjy2,ty);
 864             fjz2             = _mm_add_ps(fjz2,tz);
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             r13              = _mm_mul_ps(rsq13,rinv13);
 871             r13              = _mm_andnot_ps(dummy_mask,r13);
 872
 873             /* EWALD ELECTROSTATICS */
 874
 875             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 876             ewrt             = _mm_mul_ps(r13,ewtabscale);
 877             ewitab           = _mm_cvttps_epi32(ewrt);
 878             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 879             ewitab           = _mm_slli_epi32(ewitab,2);
 880             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 881             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 882             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 883             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 884             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 885             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 886             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 887             velec            = _mm_mul_ps(qq13,_mm_sub_ps(rinv13,velec));
 888             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
 889
 890             /* Update potential sum for this i atom from the interaction with this j atom. */
 891             velec            = _mm_andnot_ps(dummy_mask,velec);
 892             velecsum         = _mm_add_ps(velecsum,velec);
 893
 894             fscal            = felec;
 895
 896             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 897
 898             /* Calculate temporary vectorial force */
 899             tx               = _mm_mul_ps(fscal,dx13);
 900             ty               = _mm_mul_ps(fscal,dy13);
 901             tz               = _mm_mul_ps(fscal,dz13);
 902
 903             /* Update vectorial force */
 904             fix1             = _mm_add_ps(fix1,tx);
 905             fiy1             = _mm_add_ps(fiy1,ty);
 906             fiz1             = _mm_add_ps(fiz1,tz);
 907
 908             fjx3             = _mm_add_ps(fjx3,tx);
 909             fjy3             = _mm_add_ps(fjy3,ty);
 910             fjz3             = _mm_add_ps(fjz3,tz);
 911
 912             /**************************
 913              * CALCULATE INTERACTIONS *
 914              **************************/
 915
 916             r21              = _mm_mul_ps(rsq21,rinv21);
 917             r21              = _mm_andnot_ps(dummy_mask,r21);
 918
 919             /* EWALD ELECTROSTATICS */
 920
 921             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 922             ewrt             = _mm_mul_ps(r21,ewtabscale);
 923             ewitab           = _mm_cvttps_epi32(ewrt);
 924             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 925             ewitab           = _mm_slli_epi32(ewitab,2);
 926             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 927             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 928             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 929             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 930             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 931             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 932             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 933             velec            = _mm_mul_ps(qq21,_mm_sub_ps(rinv21,velec));
 934             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
 935
 936             /* Update potential sum for this i atom from the interaction with this j atom. */
 937             velec            = _mm_andnot_ps(dummy_mask,velec);
 938             velecsum         = _mm_add_ps(velecsum,velec);
 939
 940             fscal            = felec;
 941
 942             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 943
 944             /* Calculate temporary vectorial force */
 945             tx               = _mm_mul_ps(fscal,dx21);
 946             ty               = _mm_mul_ps(fscal,dy21);
 947             tz               = _mm_mul_ps(fscal,dz21);
 948
 949             /* Update vectorial force */
 950             fix2             = _mm_add_ps(fix2,tx);
 951             fiy2             = _mm_add_ps(fiy2,ty);
 952             fiz2             = _mm_add_ps(fiz2,tz);
 953
 954             fjx1             = _mm_add_ps(fjx1,tx);
 955             fjy1             = _mm_add_ps(fjy1,ty);
 956             fjz1             = _mm_add_ps(fjz1,tz);
 957
 958             /**************************
 959              * CALCULATE INTERACTIONS *
 960              **************************/
 961
 962             r22              = _mm_mul_ps(rsq22,rinv22);
 963             r22              = _mm_andnot_ps(dummy_mask,r22);
 964
 965             /* EWALD ELECTROSTATICS */
 966
 967             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 968             ewrt             = _mm_mul_ps(r22,ewtabscale);
 969             ewitab           = _mm_cvttps_epi32(ewrt);
 970             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 971             ewitab           = _mm_slli_epi32(ewitab,2);
 972             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 973             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 974             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 975             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 976             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 977             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 978             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 979             velec            = _mm_mul_ps(qq22,_mm_sub_ps(rinv22,velec));
 980             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
 981
 982             /* Update potential sum for this i atom from the interaction with this j atom. */
 983             velec            = _mm_andnot_ps(dummy_mask,velec);
 984             velecsum         = _mm_add_ps(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 989
 990             /* Calculate temporary vectorial force */
 991             tx               = _mm_mul_ps(fscal,dx22);
 992             ty               = _mm_mul_ps(fscal,dy22);
 993             tz               = _mm_mul_ps(fscal,dz22);
 994
 995             /* Update vectorial force */
 996             fix2             = _mm_add_ps(fix2,tx);
 997             fiy2             = _mm_add_ps(fiy2,ty);
 998             fiz2             = _mm_add_ps(fiz2,tz);
 999
1000             fjx2             = _mm_add_ps(fjx2,tx);
1001             fjy2             = _mm_add_ps(fjy2,ty);
1002             fjz2             = _mm_add_ps(fjz2,tz);
1003
1004             /**************************
1005              * CALCULATE INTERACTIONS *
1006              **************************/
1007
1008             r23              = _mm_mul_ps(rsq23,rinv23);
1009             r23              = _mm_andnot_ps(dummy_mask,r23);
1010
1011             /* EWALD ELECTROSTATICS */
1012
1013             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1014             ewrt             = _mm_mul_ps(r23,ewtabscale);
1015             ewitab           = _mm_cvttps_epi32(ewrt);
1016             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1017             ewitab           = _mm_slli_epi32(ewitab,2);
1018             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1019             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1020             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1021             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1022             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1023             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1024             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1025             velec            = _mm_mul_ps(qq23,_mm_sub_ps(rinv23,velec));
1026             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1027
1028             /* Update potential sum for this i atom from the interaction with this j atom. */
1029             velec            = _mm_andnot_ps(dummy_mask,velec);
1030             velecsum         = _mm_add_ps(velecsum,velec);
1031
1032             fscal            = felec;
1033
1034             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1035
1036             /* Calculate temporary vectorial force */
1037             tx               = _mm_mul_ps(fscal,dx23);
1038             ty               = _mm_mul_ps(fscal,dy23);
1039             tz               = _mm_mul_ps(fscal,dz23);
1040
1041             /* Update vectorial force */
1042             fix2             = _mm_add_ps(fix2,tx);
1043             fiy2             = _mm_add_ps(fiy2,ty);
1044             fiz2             = _mm_add_ps(fiz2,tz);
1045
1046             fjx3             = _mm_add_ps(fjx3,tx);
1047             fjy3             = _mm_add_ps(fjy3,ty);
1048             fjz3             = _mm_add_ps(fjz3,tz);
1049
1050             /**************************
1051              * CALCULATE INTERACTIONS *
1052              **************************/
1053
1054             r31              = _mm_mul_ps(rsq31,rinv31);
1055             r31              = _mm_andnot_ps(dummy_mask,r31);
1056
1057             /* EWALD ELECTROSTATICS */
1058
1059             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1060             ewrt             = _mm_mul_ps(r31,ewtabscale);
1061             ewitab           = _mm_cvttps_epi32(ewrt);
1062             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1063             ewitab           = _mm_slli_epi32(ewitab,2);
1064             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1065             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1066             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1067             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1068             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1069             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1070             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1071             velec            = _mm_mul_ps(qq31,_mm_sub_ps(rinv31,velec));
1072             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1073
1074             /* Update potential sum for this i atom from the interaction with this j atom. */
1075             velec            = _mm_andnot_ps(dummy_mask,velec);
1076             velecsum         = _mm_add_ps(velecsum,velec);
1077
1078             fscal            = felec;
1079
1080             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1081
1082             /* Calculate temporary vectorial force */
1083             tx               = _mm_mul_ps(fscal,dx31);
1084             ty               = _mm_mul_ps(fscal,dy31);
1085             tz               = _mm_mul_ps(fscal,dz31);
1086
1087             /* Update vectorial force */
1088             fix3             = _mm_add_ps(fix3,tx);
1089             fiy3             = _mm_add_ps(fiy3,ty);
1090             fiz3             = _mm_add_ps(fiz3,tz);
1091
1092             fjx1             = _mm_add_ps(fjx1,tx);
1093             fjy1             = _mm_add_ps(fjy1,ty);
1094             fjz1             = _mm_add_ps(fjz1,tz);
1095
1096             /**************************
1097              * CALCULATE INTERACTIONS *
1098              **************************/
1099
1100             r32              = _mm_mul_ps(rsq32,rinv32);
1101             r32              = _mm_andnot_ps(dummy_mask,r32);
1102
1103             /* EWALD ELECTROSTATICS */
1104
1105             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1106             ewrt             = _mm_mul_ps(r32,ewtabscale);
1107             ewitab           = _mm_cvttps_epi32(ewrt);
1108             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1109             ewitab           = _mm_slli_epi32(ewitab,2);
1110             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1111             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1112             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1113             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1114             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1115             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1116             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1117             velec            = _mm_mul_ps(qq32,_mm_sub_ps(rinv32,velec));
1118             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1119
1120             /* Update potential sum for this i atom from the interaction with this j atom. */
1121             velec            = _mm_andnot_ps(dummy_mask,velec);
1122             velecsum         = _mm_add_ps(velecsum,velec);
1123
1124             fscal            = felec;
1125
1126             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1127
1128             /* Calculate temporary vectorial force */
1129             tx               = _mm_mul_ps(fscal,dx32);
1130             ty               = _mm_mul_ps(fscal,dy32);
1131             tz               = _mm_mul_ps(fscal,dz32);
1132
1133             /* Update vectorial force */
1134             fix3             = _mm_add_ps(fix3,tx);
1135             fiy3             = _mm_add_ps(fiy3,ty);
1136             fiz3             = _mm_add_ps(fiz3,tz);
1137
1138             fjx2             = _mm_add_ps(fjx2,tx);
1139             fjy2             = _mm_add_ps(fjy2,ty);
1140             fjz2             = _mm_add_ps(fjz2,tz);
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             r33              = _mm_mul_ps(rsq33,rinv33);
1147             r33              = _mm_andnot_ps(dummy_mask,r33);
1148
1149             /* EWALD ELECTROSTATICS */
1150
1151             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1152             ewrt             = _mm_mul_ps(r33,ewtabscale);
1153             ewitab           = _mm_cvttps_epi32(ewrt);
1154             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1155             ewitab           = _mm_slli_epi32(ewitab,2);
1156             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1157             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1158             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1159             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1160             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1161             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1162             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1163             velec            = _mm_mul_ps(qq33,_mm_sub_ps(rinv33,velec));
1164             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1165
1166             /* Update potential sum for this i atom from the interaction with this j atom. */
1167             velec            = _mm_andnot_ps(dummy_mask,velec);
1168             velecsum         = _mm_add_ps(velecsum,velec);
1169
1170             fscal            = felec;
1171
1172             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1173
1174             /* Calculate temporary vectorial force */
1175             tx               = _mm_mul_ps(fscal,dx33);
1176             ty               = _mm_mul_ps(fscal,dy33);
1177             tz               = _mm_mul_ps(fscal,dz33);
1178
1179             /* Update vectorial force */
1180             fix3             = _mm_add_ps(fix3,tx);
1181             fiy3             = _mm_add_ps(fiy3,ty);
1182             fiz3             = _mm_add_ps(fiz3,tz);
1183
1184             fjx3             = _mm_add_ps(fjx3,tx);
1185             fjy3             = _mm_add_ps(fjy3,ty);
1186             fjz3             = _mm_add_ps(fjz3,tz);
1187
1188             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1189             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1190             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1191             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1192
1193             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1194                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1195
1196             /* Inner loop uses 378 flops */
1197         }
1198
1199         /* End of innermost loop */
1200
1201         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1202                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1203
1204         ggid                        = gid[iidx];
1205         /* Update potential energies */
1206         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1207
1208         /* Increment number of inner iterations */
1209         inneriter                  += j_index_end - j_index_start;
1210
1211         /* Outer loop uses 19 flops */
1212     }
1213
1214     /* Increment number of outer iterations */
1215     outeriter        += nri;
1216
1217     /* Update outer/inner flops */
1218
1219     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*378);
1220 }
1221 /*
1222  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_F_sse2_single
1223  * Electrostatics interaction: Ewald
1224  * VdW interaction:            None
1225  * Geometry:                   Water4-Water4
1226  * Calculate force/pot:        Force
1227  */
1228 void
1229 nb_kernel_ElecEw_VdwNone_GeomW4W4_F_sse2_single
1230                     (t_nblist                    * gmx_restrict       nlist,
1231                      rvec                        * gmx_restrict          xx,
1232                      rvec                        * gmx_restrict          ff,
1233                      t_forcerec                  * gmx_restrict          fr,
1234                      t_mdatoms                   * gmx_restrict     mdatoms,
1235                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1236                      t_nrnb                      * gmx_restrict        nrnb)
1237 {
1238     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1239      * just 0 for non-waters.
1240      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1241      * jnr indices corresponding to data put in the four positions in the SIMD register.
1242      */
1243     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1244     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1245     int              jnrA,jnrB,jnrC,jnrD;
1246     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1247     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1248     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1249     real             rcutoff_scalar;
1250     real             *shiftvec,*fshift,*x,*f;
1251     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1252     real             scratch[4*DIM];
1253     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1254     int              vdwioffset1;
1255     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1256     int              vdwioffset2;
1257     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1258     int              vdwioffset3;
1259     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1260     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1261     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1262     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1263     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1264     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1265     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1266     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1267     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1268     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1269     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1270     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1271     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1272     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1273     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1274     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1275     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1276     real             *charge;
1277     __m128i          ewitab;
1278     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1279     real             *ewtab;
1280     __m128           dummy_mask,cutoff_mask;
1281     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1282     __m128           one     = _mm_set1_ps(1.0);
1283     __m128           two     = _mm_set1_ps(2.0);
1284     x                = xx[0];
1285     f                = ff[0];
1286
1287     nri              = nlist->nri;
1288     iinr             = nlist->iinr;
1289     jindex           = nlist->jindex;
1290     jjnr             = nlist->jjnr;
1291     shiftidx         = nlist->shift;
1292     gid              = nlist->gid;
1293     shiftvec         = fr->shift_vec[0];
1294     fshift           = fr->fshift[0];
1295     facel            = _mm_set1_ps(fr->epsfac);
1296     charge           = mdatoms->chargeA;
1297
1298     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1299     ewtab            = fr->ic->tabq_coul_F;
1300     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1301     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1302
1303     /* Setup water-specific parameters */
1304     inr              = nlist->iinr[0];
1305     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1306     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1307     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1308
1309     jq1              = _mm_set1_ps(charge[inr+1]);
1310     jq2              = _mm_set1_ps(charge[inr+2]);
1311     jq3              = _mm_set1_ps(charge[inr+3]);
1312     qq11             = _mm_mul_ps(iq1,jq1);
1313     qq12             = _mm_mul_ps(iq1,jq2);
1314     qq13             = _mm_mul_ps(iq1,jq3);
1315     qq21             = _mm_mul_ps(iq2,jq1);
1316     qq22             = _mm_mul_ps(iq2,jq2);
1317     qq23             = _mm_mul_ps(iq2,jq3);
1318     qq31             = _mm_mul_ps(iq3,jq1);
1319     qq32             = _mm_mul_ps(iq3,jq2);
1320     qq33             = _mm_mul_ps(iq3,jq3);
1321
1322     /* Avoid stupid compiler warnings */
1323     jnrA = jnrB = jnrC = jnrD = 0;
1324     j_coord_offsetA = 0;
1325     j_coord_offsetB = 0;
1326     j_coord_offsetC = 0;
1327     j_coord_offsetD = 0;
1328
1329     outeriter        = 0;
1330     inneriter        = 0;
1331
1332     for(iidx=0;iidx<4*DIM;iidx++)
1333     {
1334         scratch[iidx] = 0.0;
1335     }
1336
1337     /* Start outer loop over neighborlists */
1338     for(iidx=0; iidx<nri; iidx++)
1339     {
1340         /* Load shift vector for this list */
1341         i_shift_offset   = DIM*shiftidx[iidx];
1342
1343         /* Load limits for loop over neighbors */
1344         j_index_start    = jindex[iidx];
1345         j_index_end      = jindex[iidx+1];
1346
1347         /* Get outer coordinate index */
1348         inr              = iinr[iidx];
1349         i_coord_offset   = DIM*inr;
1350
1351         /* Load i particle coords and add shift vector */
1352         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1353                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1354
1355         fix1             = _mm_setzero_ps();
1356         fiy1             = _mm_setzero_ps();
1357         fiz1             = _mm_setzero_ps();
1358         fix2             = _mm_setzero_ps();
1359         fiy2             = _mm_setzero_ps();
1360         fiz2             = _mm_setzero_ps();
1361         fix3             = _mm_setzero_ps();
1362         fiy3             = _mm_setzero_ps();
1363         fiz3             = _mm_setzero_ps();
1364
1365         /* Start inner kernel loop */
1366         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1367         {
1368
1369             /* Get j neighbor index, and coordinate index */
1370             jnrA             = jjnr[jidx];
1371             jnrB             = jjnr[jidx+1];
1372             jnrC             = jjnr[jidx+2];
1373             jnrD             = jjnr[jidx+3];
1374             j_coord_offsetA  = DIM*jnrA;
1375             j_coord_offsetB  = DIM*jnrB;
1376             j_coord_offsetC  = DIM*jnrC;
1377             j_coord_offsetD  = DIM*jnrD;
1378
1379             /* load j atom coordinates */
1380             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1381                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1382                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1383
1384             /* Calculate displacement vector */
1385             dx11             = _mm_sub_ps(ix1,jx1);
1386             dy11             = _mm_sub_ps(iy1,jy1);
1387             dz11             = _mm_sub_ps(iz1,jz1);
1388             dx12             = _mm_sub_ps(ix1,jx2);
1389             dy12             = _mm_sub_ps(iy1,jy2);
1390             dz12             = _mm_sub_ps(iz1,jz2);
1391             dx13             = _mm_sub_ps(ix1,jx3);
1392             dy13             = _mm_sub_ps(iy1,jy3);
1393             dz13             = _mm_sub_ps(iz1,jz3);
1394             dx21             = _mm_sub_ps(ix2,jx1);
1395             dy21             = _mm_sub_ps(iy2,jy1);
1396             dz21             = _mm_sub_ps(iz2,jz1);
1397             dx22             = _mm_sub_ps(ix2,jx2);
1398             dy22             = _mm_sub_ps(iy2,jy2);
1399             dz22             = _mm_sub_ps(iz2,jz2);
1400             dx23             = _mm_sub_ps(ix2,jx3);
1401             dy23             = _mm_sub_ps(iy2,jy3);
1402             dz23             = _mm_sub_ps(iz2,jz3);
1403             dx31             = _mm_sub_ps(ix3,jx1);
1404             dy31             = _mm_sub_ps(iy3,jy1);
1405             dz31             = _mm_sub_ps(iz3,jz1);
1406             dx32             = _mm_sub_ps(ix3,jx2);
1407             dy32             = _mm_sub_ps(iy3,jy2);
1408             dz32             = _mm_sub_ps(iz3,jz2);
1409             dx33             = _mm_sub_ps(ix3,jx3);
1410             dy33             = _mm_sub_ps(iy3,jy3);
1411             dz33             = _mm_sub_ps(iz3,jz3);
1412
1413             /* Calculate squared distance and things based on it */
1414             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1415             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1416             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1417             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1418             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1419             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1420             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1421             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1422             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1423
1424             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1425             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1426             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1427             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1428             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1429             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1430             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1431             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1432             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1433
1434             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1435             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1436             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1437             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1438             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1439             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1440             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1441             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1442             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1443
1444             fjx1             = _mm_setzero_ps();
1445             fjy1             = _mm_setzero_ps();
1446             fjz1             = _mm_setzero_ps();
1447             fjx2             = _mm_setzero_ps();
1448             fjy2             = _mm_setzero_ps();
1449             fjz2             = _mm_setzero_ps();
1450             fjx3             = _mm_setzero_ps();
1451             fjy3             = _mm_setzero_ps();
1452             fjz3             = _mm_setzero_ps();
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             r11              = _mm_mul_ps(rsq11,rinv11);
1459
1460             /* EWALD ELECTROSTATICS */
1461
1462             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1463             ewrt             = _mm_mul_ps(r11,ewtabscale);
1464             ewitab           = _mm_cvttps_epi32(ewrt);
1465             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1466             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1467                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1468                                          &ewtabF,&ewtabFn);
1469             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1470             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1471
1472             fscal            = felec;
1473
1474             /* Calculate temporary vectorial force */
1475             tx               = _mm_mul_ps(fscal,dx11);
1476             ty               = _mm_mul_ps(fscal,dy11);
1477             tz               = _mm_mul_ps(fscal,dz11);
1478
1479             /* Update vectorial force */
1480             fix1             = _mm_add_ps(fix1,tx);
1481             fiy1             = _mm_add_ps(fiy1,ty);
1482             fiz1             = _mm_add_ps(fiz1,tz);
1483
1484             fjx1             = _mm_add_ps(fjx1,tx);
1485             fjy1             = _mm_add_ps(fjy1,ty);
1486             fjz1             = _mm_add_ps(fjz1,tz);
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             r12              = _mm_mul_ps(rsq12,rinv12);
1493
1494             /* EWALD ELECTROSTATICS */
1495
1496             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1497             ewrt             = _mm_mul_ps(r12,ewtabscale);
1498             ewitab           = _mm_cvttps_epi32(ewrt);
1499             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1500             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1501                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1502                                          &ewtabF,&ewtabFn);
1503             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1504             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1505
1506             fscal            = felec;
1507
1508             /* Calculate temporary vectorial force */
1509             tx               = _mm_mul_ps(fscal,dx12);
1510             ty               = _mm_mul_ps(fscal,dy12);
1511             tz               = _mm_mul_ps(fscal,dz12);
1512
1513             /* Update vectorial force */
1514             fix1             = _mm_add_ps(fix1,tx);
1515             fiy1             = _mm_add_ps(fiy1,ty);
1516             fiz1             = _mm_add_ps(fiz1,tz);
1517
1518             fjx2             = _mm_add_ps(fjx2,tx);
1519             fjy2             = _mm_add_ps(fjy2,ty);
1520             fjz2             = _mm_add_ps(fjz2,tz);
1521
1522             /**************************
1523              * CALCULATE INTERACTIONS *
1524              **************************/
1525
1526             r13              = _mm_mul_ps(rsq13,rinv13);
1527
1528             /* EWALD ELECTROSTATICS */
1529
1530             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1531             ewrt             = _mm_mul_ps(r13,ewtabscale);
1532             ewitab           = _mm_cvttps_epi32(ewrt);
1533             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1534             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1535                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1536                                          &ewtabF,&ewtabFn);
1537             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1538             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1539
1540             fscal            = felec;
1541
1542             /* Calculate temporary vectorial force */
1543             tx               = _mm_mul_ps(fscal,dx13);
1544             ty               = _mm_mul_ps(fscal,dy13);
1545             tz               = _mm_mul_ps(fscal,dz13);
1546
1547             /* Update vectorial force */
1548             fix1             = _mm_add_ps(fix1,tx);
1549             fiy1             = _mm_add_ps(fiy1,ty);
1550             fiz1             = _mm_add_ps(fiz1,tz);
1551
1552             fjx3             = _mm_add_ps(fjx3,tx);
1553             fjy3             = _mm_add_ps(fjy3,ty);
1554             fjz3             = _mm_add_ps(fjz3,tz);
1555
1556             /**************************
1557              * CALCULATE INTERACTIONS *
1558              **************************/
1559
1560             r21              = _mm_mul_ps(rsq21,rinv21);
1561
1562             /* EWALD ELECTROSTATICS */
1563
1564             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1565             ewrt             = _mm_mul_ps(r21,ewtabscale);
1566             ewitab           = _mm_cvttps_epi32(ewrt);
1567             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1568             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1569                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1570                                          &ewtabF,&ewtabFn);
1571             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1572             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1573
1574             fscal            = felec;
1575
1576             /* Calculate temporary vectorial force */
1577             tx               = _mm_mul_ps(fscal,dx21);
1578             ty               = _mm_mul_ps(fscal,dy21);
1579             tz               = _mm_mul_ps(fscal,dz21);
1580
1581             /* Update vectorial force */
1582             fix2             = _mm_add_ps(fix2,tx);
1583             fiy2             = _mm_add_ps(fiy2,ty);
1584             fiz2             = _mm_add_ps(fiz2,tz);
1585
1586             fjx1             = _mm_add_ps(fjx1,tx);
1587             fjy1             = _mm_add_ps(fjy1,ty);
1588             fjz1             = _mm_add_ps(fjz1,tz);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             r22              = _mm_mul_ps(rsq22,rinv22);
1595
1596             /* EWALD ELECTROSTATICS */
1597
1598             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1599             ewrt             = _mm_mul_ps(r22,ewtabscale);
1600             ewitab           = _mm_cvttps_epi32(ewrt);
1601             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1602             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1603                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1604                                          &ewtabF,&ewtabFn);
1605             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1606             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
1607
1608             fscal            = felec;
1609
1610             /* Calculate temporary vectorial force */
1611             tx               = _mm_mul_ps(fscal,dx22);
1612             ty               = _mm_mul_ps(fscal,dy22);
1613             tz               = _mm_mul_ps(fscal,dz22);
1614
1615             /* Update vectorial force */
1616             fix2             = _mm_add_ps(fix2,tx);
1617             fiy2             = _mm_add_ps(fiy2,ty);
1618             fiz2             = _mm_add_ps(fiz2,tz);
1619
1620             fjx2             = _mm_add_ps(fjx2,tx);
1621             fjy2             = _mm_add_ps(fjy2,ty);
1622             fjz2             = _mm_add_ps(fjz2,tz);
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             r23              = _mm_mul_ps(rsq23,rinv23);
1629
1630             /* EWALD ELECTROSTATICS */
1631
1632             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1633             ewrt             = _mm_mul_ps(r23,ewtabscale);
1634             ewitab           = _mm_cvttps_epi32(ewrt);
1635             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1636             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1637                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1638                                          &ewtabF,&ewtabFn);
1639             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1640             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
1641
1642             fscal            = felec;
1643
1644             /* Calculate temporary vectorial force */
1645             tx               = _mm_mul_ps(fscal,dx23);
1646             ty               = _mm_mul_ps(fscal,dy23);
1647             tz               = _mm_mul_ps(fscal,dz23);
1648
1649             /* Update vectorial force */
1650             fix2             = _mm_add_ps(fix2,tx);
1651             fiy2             = _mm_add_ps(fiy2,ty);
1652             fiz2             = _mm_add_ps(fiz2,tz);
1653
1654             fjx3             = _mm_add_ps(fjx3,tx);
1655             fjy3             = _mm_add_ps(fjy3,ty);
1656             fjz3             = _mm_add_ps(fjz3,tz);
1657
1658             /**************************
1659              * CALCULATE INTERACTIONS *
1660              **************************/
1661
1662             r31              = _mm_mul_ps(rsq31,rinv31);
1663
1664             /* EWALD ELECTROSTATICS */
1665
1666             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1667             ewrt             = _mm_mul_ps(r31,ewtabscale);
1668             ewitab           = _mm_cvttps_epi32(ewrt);
1669             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1670             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1671                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1672                                          &ewtabF,&ewtabFn);
1673             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1674             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
1675
1676             fscal            = felec;
1677
1678             /* Calculate temporary vectorial force */
1679             tx               = _mm_mul_ps(fscal,dx31);
1680             ty               = _mm_mul_ps(fscal,dy31);
1681             tz               = _mm_mul_ps(fscal,dz31);
1682
1683             /* Update vectorial force */
1684             fix3             = _mm_add_ps(fix3,tx);
1685             fiy3             = _mm_add_ps(fiy3,ty);
1686             fiz3             = _mm_add_ps(fiz3,tz);
1687
1688             fjx1             = _mm_add_ps(fjx1,tx);
1689             fjy1             = _mm_add_ps(fjy1,ty);
1690             fjz1             = _mm_add_ps(fjz1,tz);
1691
1692             /**************************
1693              * CALCULATE INTERACTIONS *
1694              **************************/
1695
1696             r32              = _mm_mul_ps(rsq32,rinv32);
1697
1698             /* EWALD ELECTROSTATICS */
1699
1700             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1701             ewrt             = _mm_mul_ps(r32,ewtabscale);
1702             ewitab           = _mm_cvttps_epi32(ewrt);
1703             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1704             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1705                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1706                                          &ewtabF,&ewtabFn);
1707             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1708             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
1709
1710             fscal            = felec;
1711
1712             /* Calculate temporary vectorial force */
1713             tx               = _mm_mul_ps(fscal,dx32);
1714             ty               = _mm_mul_ps(fscal,dy32);
1715             tz               = _mm_mul_ps(fscal,dz32);
1716
1717             /* Update vectorial force */
1718             fix3             = _mm_add_ps(fix3,tx);
1719             fiy3             = _mm_add_ps(fiy3,ty);
1720             fiz3             = _mm_add_ps(fiz3,tz);
1721
1722             fjx2             = _mm_add_ps(fjx2,tx);
1723             fjy2             = _mm_add_ps(fjy2,ty);
1724             fjz2             = _mm_add_ps(fjz2,tz);
1725
1726             /**************************
1727              * CALCULATE INTERACTIONS *
1728              **************************/
1729
1730             r33              = _mm_mul_ps(rsq33,rinv33);
1731
1732             /* EWALD ELECTROSTATICS */
1733
1734             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1735             ewrt             = _mm_mul_ps(r33,ewtabscale);
1736             ewitab           = _mm_cvttps_epi32(ewrt);
1737             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1738             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1739                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1740                                          &ewtabF,&ewtabFn);
1741             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1742             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
1743
1744             fscal            = felec;
1745
1746             /* Calculate temporary vectorial force */
1747             tx               = _mm_mul_ps(fscal,dx33);
1748             ty               = _mm_mul_ps(fscal,dy33);
1749             tz               = _mm_mul_ps(fscal,dz33);
1750
1751             /* Update vectorial force */
1752             fix3             = _mm_add_ps(fix3,tx);
1753             fiy3             = _mm_add_ps(fiy3,ty);
1754             fiz3             = _mm_add_ps(fiz3,tz);
1755
1756             fjx3             = _mm_add_ps(fjx3,tx);
1757             fjy3             = _mm_add_ps(fjy3,ty);
1758             fjz3             = _mm_add_ps(fjz3,tz);
1759
1760             fjptrA             = f+j_coord_offsetA;
1761             fjptrB             = f+j_coord_offsetB;
1762             fjptrC             = f+j_coord_offsetC;
1763             fjptrD             = f+j_coord_offsetD;
1764
1765             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1766                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1767
1768             /* Inner loop uses 324 flops */
1769         }
1770
1771         if(jidx<j_index_end)
1772         {
1773
1774             /* Get j neighbor index, and coordinate index */
1775             jnrlistA         = jjnr[jidx];
1776             jnrlistB         = jjnr[jidx+1];
1777             jnrlistC         = jjnr[jidx+2];
1778             jnrlistD         = jjnr[jidx+3];
1779             /* Sign of each element will be negative for non-real atoms.
1780              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1781              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1782              */
1783             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1784             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1785             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1786             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1787             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1788             j_coord_offsetA  = DIM*jnrA;
1789             j_coord_offsetB  = DIM*jnrB;
1790             j_coord_offsetC  = DIM*jnrC;
1791             j_coord_offsetD  = DIM*jnrD;
1792
1793             /* load j atom coordinates */
1794             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1795                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1796                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1797
1798             /* Calculate displacement vector */
1799             dx11             = _mm_sub_ps(ix1,jx1);
1800             dy11             = _mm_sub_ps(iy1,jy1);
1801             dz11             = _mm_sub_ps(iz1,jz1);
1802             dx12             = _mm_sub_ps(ix1,jx2);
1803             dy12             = _mm_sub_ps(iy1,jy2);
1804             dz12             = _mm_sub_ps(iz1,jz2);
1805             dx13             = _mm_sub_ps(ix1,jx3);
1806             dy13             = _mm_sub_ps(iy1,jy3);
1807             dz13             = _mm_sub_ps(iz1,jz3);
1808             dx21             = _mm_sub_ps(ix2,jx1);
1809             dy21             = _mm_sub_ps(iy2,jy1);
1810             dz21             = _mm_sub_ps(iz2,jz1);
1811             dx22             = _mm_sub_ps(ix2,jx2);
1812             dy22             = _mm_sub_ps(iy2,jy2);
1813             dz22             = _mm_sub_ps(iz2,jz2);
1814             dx23             = _mm_sub_ps(ix2,jx3);
1815             dy23             = _mm_sub_ps(iy2,jy3);
1816             dz23             = _mm_sub_ps(iz2,jz3);
1817             dx31             = _mm_sub_ps(ix3,jx1);
1818             dy31             = _mm_sub_ps(iy3,jy1);
1819             dz31             = _mm_sub_ps(iz3,jz1);
1820             dx32             = _mm_sub_ps(ix3,jx2);
1821             dy32             = _mm_sub_ps(iy3,jy2);
1822             dz32             = _mm_sub_ps(iz3,jz2);
1823             dx33             = _mm_sub_ps(ix3,jx3);
1824             dy33             = _mm_sub_ps(iy3,jy3);
1825             dz33             = _mm_sub_ps(iz3,jz3);
1826
1827             /* Calculate squared distance and things based on it */
1828             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1829             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1830             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1831             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1832             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1833             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1834             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1835             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1836             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1837
1838             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1839             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1840             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1841             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1842             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1843             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1844             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1845             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1846             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1847
1848             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1849             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1850             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1851             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1852             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1853             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1854             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1855             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1856             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1857
1858             fjx1             = _mm_setzero_ps();
1859             fjy1             = _mm_setzero_ps();
1860             fjz1             = _mm_setzero_ps();
1861             fjx2             = _mm_setzero_ps();
1862             fjy2             = _mm_setzero_ps();
1863             fjz2             = _mm_setzero_ps();
1864             fjx3             = _mm_setzero_ps();
1865             fjy3             = _mm_setzero_ps();
1866             fjz3             = _mm_setzero_ps();
1867
1868             /**************************
1869              * CALCULATE INTERACTIONS *
1870              **************************/
1871
1872             r11              = _mm_mul_ps(rsq11,rinv11);
1873             r11              = _mm_andnot_ps(dummy_mask,r11);
1874
1875             /* EWALD ELECTROSTATICS */
1876
1877             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1878             ewrt             = _mm_mul_ps(r11,ewtabscale);
1879             ewitab           = _mm_cvttps_epi32(ewrt);
1880             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1881             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1882                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1883                                          &ewtabF,&ewtabFn);
1884             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1885             felec            = _mm_mul_ps(_mm_mul_ps(qq11,rinv11),_mm_sub_ps(rinvsq11,felec));
1886
1887             fscal            = felec;
1888
1889             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1890
1891             /* Calculate temporary vectorial force */
1892             tx               = _mm_mul_ps(fscal,dx11);
1893             ty               = _mm_mul_ps(fscal,dy11);
1894             tz               = _mm_mul_ps(fscal,dz11);
1895
1896             /* Update vectorial force */
1897             fix1             = _mm_add_ps(fix1,tx);
1898             fiy1             = _mm_add_ps(fiy1,ty);
1899             fiz1             = _mm_add_ps(fiz1,tz);
1900
1901             fjx1             = _mm_add_ps(fjx1,tx);
1902             fjy1             = _mm_add_ps(fjy1,ty);
1903             fjz1             = _mm_add_ps(fjz1,tz);
1904
1905             /**************************
1906              * CALCULATE INTERACTIONS *
1907              **************************/
1908
1909             r12              = _mm_mul_ps(rsq12,rinv12);
1910             r12              = _mm_andnot_ps(dummy_mask,r12);
1911
1912             /* EWALD ELECTROSTATICS */
1913
1914             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1915             ewrt             = _mm_mul_ps(r12,ewtabscale);
1916             ewitab           = _mm_cvttps_epi32(ewrt);
1917             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1918             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1919                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1920                                          &ewtabF,&ewtabFn);
1921             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1922             felec            = _mm_mul_ps(_mm_mul_ps(qq12,rinv12),_mm_sub_ps(rinvsq12,felec));
1923
1924             fscal            = felec;
1925
1926             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1927
1928             /* Calculate temporary vectorial force */
1929             tx               = _mm_mul_ps(fscal,dx12);
1930             ty               = _mm_mul_ps(fscal,dy12);
1931             tz               = _mm_mul_ps(fscal,dz12);
1932
1933             /* Update vectorial force */
1934             fix1             = _mm_add_ps(fix1,tx);
1935             fiy1             = _mm_add_ps(fiy1,ty);
1936             fiz1             = _mm_add_ps(fiz1,tz);
1937
1938             fjx2             = _mm_add_ps(fjx2,tx);
1939             fjy2             = _mm_add_ps(fjy2,ty);
1940             fjz2             = _mm_add_ps(fjz2,tz);
1941
1942             /**************************
1943              * CALCULATE INTERACTIONS *
1944              **************************/
1945
1946             r13              = _mm_mul_ps(rsq13,rinv13);
1947             r13              = _mm_andnot_ps(dummy_mask,r13);
1948
1949             /* EWALD ELECTROSTATICS */
1950
1951             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1952             ewrt             = _mm_mul_ps(r13,ewtabscale);
1953             ewitab           = _mm_cvttps_epi32(ewrt);
1954             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1955             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1956                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1957                                          &ewtabF,&ewtabFn);
1958             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1959             felec            = _mm_mul_ps(_mm_mul_ps(qq13,rinv13),_mm_sub_ps(rinvsq13,felec));
1960
1961             fscal            = felec;
1962
1963             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1964
1965             /* Calculate temporary vectorial force */
1966             tx               = _mm_mul_ps(fscal,dx13);
1967             ty               = _mm_mul_ps(fscal,dy13);
1968             tz               = _mm_mul_ps(fscal,dz13);
1969
1970             /* Update vectorial force */
1971             fix1             = _mm_add_ps(fix1,tx);
1972             fiy1             = _mm_add_ps(fiy1,ty);
1973             fiz1             = _mm_add_ps(fiz1,tz);
1974
1975             fjx3             = _mm_add_ps(fjx3,tx);
1976             fjy3             = _mm_add_ps(fjy3,ty);
1977             fjz3             = _mm_add_ps(fjz3,tz);
1978
1979             /**************************
1980              * CALCULATE INTERACTIONS *
1981              **************************/
1982
1983             r21              = _mm_mul_ps(rsq21,rinv21);
1984             r21              = _mm_andnot_ps(dummy_mask,r21);
1985
1986             /* EWALD ELECTROSTATICS */
1987
1988             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1989             ewrt             = _mm_mul_ps(r21,ewtabscale);
1990             ewitab           = _mm_cvttps_epi32(ewrt);
1991             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
1992             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1993                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
1994                                          &ewtabF,&ewtabFn);
1995             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
1996             felec            = _mm_mul_ps(_mm_mul_ps(qq21,rinv21),_mm_sub_ps(rinvsq21,felec));
1997
1998             fscal            = felec;
1999
2000             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2001
2002             /* Calculate temporary vectorial force */
2003             tx               = _mm_mul_ps(fscal,dx21);
2004             ty               = _mm_mul_ps(fscal,dy21);
2005             tz               = _mm_mul_ps(fscal,dz21);
2006
2007             /* Update vectorial force */
2008             fix2             = _mm_add_ps(fix2,tx);
2009             fiy2             = _mm_add_ps(fiy2,ty);
2010             fiz2             = _mm_add_ps(fiz2,tz);
2011
2012             fjx1             = _mm_add_ps(fjx1,tx);
2013             fjy1             = _mm_add_ps(fjy1,ty);
2014             fjz1             = _mm_add_ps(fjz1,tz);
2015
2016             /**************************
2017              * CALCULATE INTERACTIONS *
2018              **************************/
2019
2020             r22              = _mm_mul_ps(rsq22,rinv22);
2021             r22              = _mm_andnot_ps(dummy_mask,r22);
2022
2023             /* EWALD ELECTROSTATICS */
2024
2025             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2026             ewrt             = _mm_mul_ps(r22,ewtabscale);
2027             ewitab           = _mm_cvttps_epi32(ewrt);
2028             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2029             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2030                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2031                                          &ewtabF,&ewtabFn);
2032             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2033             felec            = _mm_mul_ps(_mm_mul_ps(qq22,rinv22),_mm_sub_ps(rinvsq22,felec));
2034
2035             fscal            = felec;
2036
2037             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2038
2039             /* Calculate temporary vectorial force */
2040             tx               = _mm_mul_ps(fscal,dx22);
2041             ty               = _mm_mul_ps(fscal,dy22);
2042             tz               = _mm_mul_ps(fscal,dz22);
2043
2044             /* Update vectorial force */
2045             fix2             = _mm_add_ps(fix2,tx);
2046             fiy2             = _mm_add_ps(fiy2,ty);
2047             fiz2             = _mm_add_ps(fiz2,tz);
2048
2049             fjx2             = _mm_add_ps(fjx2,tx);
2050             fjy2             = _mm_add_ps(fjy2,ty);
2051             fjz2             = _mm_add_ps(fjz2,tz);
2052
2053             /**************************
2054              * CALCULATE INTERACTIONS *
2055              **************************/
2056
2057             r23              = _mm_mul_ps(rsq23,rinv23);
2058             r23              = _mm_andnot_ps(dummy_mask,r23);
2059
2060             /* EWALD ELECTROSTATICS */
2061
2062             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2063             ewrt             = _mm_mul_ps(r23,ewtabscale);
2064             ewitab           = _mm_cvttps_epi32(ewrt);
2065             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2066             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2067                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2068                                          &ewtabF,&ewtabFn);
2069             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2070             felec            = _mm_mul_ps(_mm_mul_ps(qq23,rinv23),_mm_sub_ps(rinvsq23,felec));
2071
2072             fscal            = felec;
2073
2074             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2075
2076             /* Calculate temporary vectorial force */
2077             tx               = _mm_mul_ps(fscal,dx23);
2078             ty               = _mm_mul_ps(fscal,dy23);
2079             tz               = _mm_mul_ps(fscal,dz23);
2080
2081             /* Update vectorial force */
2082             fix2             = _mm_add_ps(fix2,tx);
2083             fiy2             = _mm_add_ps(fiy2,ty);
2084             fiz2             = _mm_add_ps(fiz2,tz);
2085
2086             fjx3             = _mm_add_ps(fjx3,tx);
2087             fjy3             = _mm_add_ps(fjy3,ty);
2088             fjz3             = _mm_add_ps(fjz3,tz);
2089
2090             /**************************
2091              * CALCULATE INTERACTIONS *
2092              **************************/
2093
2094             r31              = _mm_mul_ps(rsq31,rinv31);
2095             r31              = _mm_andnot_ps(dummy_mask,r31);
2096
2097             /* EWALD ELECTROSTATICS */
2098
2099             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2100             ewrt             = _mm_mul_ps(r31,ewtabscale);
2101             ewitab           = _mm_cvttps_epi32(ewrt);
2102             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2103             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2104                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2105                                          &ewtabF,&ewtabFn);
2106             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2107             felec            = _mm_mul_ps(_mm_mul_ps(qq31,rinv31),_mm_sub_ps(rinvsq31,felec));
2108
2109             fscal            = felec;
2110
2111             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2112
2113             /* Calculate temporary vectorial force */
2114             tx               = _mm_mul_ps(fscal,dx31);
2115             ty               = _mm_mul_ps(fscal,dy31);
2116             tz               = _mm_mul_ps(fscal,dz31);
2117
2118             /* Update vectorial force */
2119             fix3             = _mm_add_ps(fix3,tx);
2120             fiy3             = _mm_add_ps(fiy3,ty);
2121             fiz3             = _mm_add_ps(fiz3,tz);
2122
2123             fjx1             = _mm_add_ps(fjx1,tx);
2124             fjy1             = _mm_add_ps(fjy1,ty);
2125             fjz1             = _mm_add_ps(fjz1,tz);
2126
2127             /**************************
2128              * CALCULATE INTERACTIONS *
2129              **************************/
2130
2131             r32              = _mm_mul_ps(rsq32,rinv32);
2132             r32              = _mm_andnot_ps(dummy_mask,r32);
2133
2134             /* EWALD ELECTROSTATICS */
2135
2136             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2137             ewrt             = _mm_mul_ps(r32,ewtabscale);
2138             ewitab           = _mm_cvttps_epi32(ewrt);
2139             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2140             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2141                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2142                                          &ewtabF,&ewtabFn);
2143             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2144             felec            = _mm_mul_ps(_mm_mul_ps(qq32,rinv32),_mm_sub_ps(rinvsq32,felec));
2145
2146             fscal            = felec;
2147
2148             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2149
2150             /* Calculate temporary vectorial force */
2151             tx               = _mm_mul_ps(fscal,dx32);
2152             ty               = _mm_mul_ps(fscal,dy32);
2153             tz               = _mm_mul_ps(fscal,dz32);
2154
2155             /* Update vectorial force */
2156             fix3             = _mm_add_ps(fix3,tx);
2157             fiy3             = _mm_add_ps(fiy3,ty);
2158             fiz3             = _mm_add_ps(fiz3,tz);
2159
2160             fjx2             = _mm_add_ps(fjx2,tx);
2161             fjy2             = _mm_add_ps(fjy2,ty);
2162             fjz2             = _mm_add_ps(fjz2,tz);
2163
2164             /**************************
2165              * CALCULATE INTERACTIONS *
2166              **************************/
2167
2168             r33              = _mm_mul_ps(rsq33,rinv33);
2169             r33              = _mm_andnot_ps(dummy_mask,r33);
2170
2171             /* EWALD ELECTROSTATICS */
2172
2173             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2174             ewrt             = _mm_mul_ps(r33,ewtabscale);
2175             ewitab           = _mm_cvttps_epi32(ewrt);
2176             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
2177             gmx_mm_load_4pair_swizzle_ps(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
2178                                          ewtab+gmx_mm_extract_epi32(ewitab,2),ewtab+gmx_mm_extract_epi32(ewitab,3),
2179                                          &ewtabF,&ewtabFn);
2180             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
2181             felec            = _mm_mul_ps(_mm_mul_ps(qq33,rinv33),_mm_sub_ps(rinvsq33,felec));
2182
2183             fscal            = felec;
2184
2185             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2186
2187             /* Calculate temporary vectorial force */
2188             tx               = _mm_mul_ps(fscal,dx33);
2189             ty               = _mm_mul_ps(fscal,dy33);
2190             tz               = _mm_mul_ps(fscal,dz33);
2191
2192             /* Update vectorial force */
2193             fix3             = _mm_add_ps(fix3,tx);
2194             fiy3             = _mm_add_ps(fiy3,ty);
2195             fiz3             = _mm_add_ps(fiz3,tz);
2196
2197             fjx3             = _mm_add_ps(fjx3,tx);
2198             fjy3             = _mm_add_ps(fjy3,ty);
2199             fjz3             = _mm_add_ps(fjz3,tz);
2200
2201             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2202             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2203             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2204             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2205
2206             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
2207                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2208
2209             /* Inner loop uses 333 flops */
2210         }
2211
2212         /* End of innermost loop */
2213
2214         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2215                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2216
2217         /* Increment number of inner iterations */
2218         inneriter                  += j_index_end - j_index_start;
2219
2220         /* Outer loop uses 18 flops */
2221     }
2222
2223     /* Increment number of outer iterations */
2224     outeriter        += nri;
2225
2226     /* Update outer/inner flops */
2227
2228     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*333);
2229 }